See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224441000

The SIGMA Algorithm: A Glottal Activity Detector for Electroglottographic

Signals

Article  in  IEEE Transactions on Audio Speech and Language Processing · December 2009

DOI: 10.1109/TASL.2009.2022430 · Source: IEEE Xplore

CITATIONS READS
77 546

2 authors, including:

Mark R. P. Thomas
Dolby Laboratories, Inc.
54 PUBLICATIONS   945 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Mark R. P. Thomas on 06 March 2015.

The user has requested enhancement of the downloaded file.

IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 8, NOVEMBER 2009 1557

The SIGMA Algorithm: A Glottal Activity Detector

for Electroglottographic Signals
Mark R. P. Thomas, Student Member, IEEE, and Patrick A. Naylor, Senior Member, IEEE

Abstract—Accurate estimation of glottal closure instants (GCIs)

and opening instants (GOIs) is important for speech processing
applications that benefit from glottal-synchronous processing.
The majority of existing approaches detect GCIs by comparing
the differentiated EGG signal to a threshold and are able to
provide accurate results during voiced speech. More recent al-
gorithms use a similar approach across multiple dyadic scales
using the stationary wavelet transform. All existing approaches
are however prone to errors around the transition regions at the
end of voiced segments of speech. This paper describes a new
method for EGG-based glottal activity detection which exhibits
high accuracy over the entirety of voiced segments, including, in
particular, the transition regions, thereby giving significant im-
provement over existing methods. Following a stationary wavelet
transform-based preprocessor, detection of excitation due to
glottal closure is performed using a group delay function and then
true and false detections are discriminated by Gaussian mixture
modeling. GOI detection involves additional processing using the
estimated GCIs. The main purpose of our algorithm is to provide
a ground-truth for GCIs and GOIs. This is essential in order to
evaluate algorithms that estimate GCIs and GOIs from the speech
signal only, and is also of high value in the analysis of pathological Fig. 1. (a) Speech signal, (b) the corresponding EGG signal, and (c) the EGG
speech where knowledge of GCIs and GOIs is often needed. We time derivative for / /. Negative peaks due to glottal opening are weak in (c).
compare our algorithm with two previous algorithms against a
hand-labeled database. Evaluation has shown an average GCI hit
rate of 99.47% and GOI of 99.35%, compared to 96.08 and 92.54 instant (GOI). The process repeats periodically as a series of
for the best-performing existing algorithm. pulses that produces “modal” voiced speech. The ratio of open
Index Terms—Electroglottograph (EGG), glottal closure in- time with respect to glottal period is termed the open quotient
stants (GCIs), group delay function, laryngograph. (OQ) [1].
Identification of GCIs in voiced speech is important for
speech processing algorithms such as pitch tracking [2],
I. INTRODUCTION prosodic speech modification [3], speech dereverberation [4],
glottal-synchronous processing in speech synthesis [5] and

A LL voiced sounds are produced by an excitation signal

that is filtered by a passive resonator called the vocal tract.
This excitation is produced by the vocal folds, which consist
of opposing ligaments that form a constriction at the top of the
trachea as it joins the lower vocal tract. When air is expelled
from the lungs at sufficient velocity through this orifice—often
referred to as the glottis—the Bernoulli Effect results in a par-
tial vacuum that causes the vocal folds to snap shut and disrupt
the air flow. This glottal closure instant (GCI) is followed by a
period during which the glottis is closed, until muscle tension
and air pressure cause the folds to reopen at the glottal opening
Manuscript received September 30, 2008; revised March 10, 2009. First pub-
lished May 08, 2009; current version published August 14, 2009. The associate
editor coordinating the review of this manuscript and approving it for publica-
tion was Dr. Hiroshi Sawada.
The authors are with the Department of Electrical and Electronic En-
gineering, Imperial College London, London SW7 2AZ, U.K. (e-mail:
mark.r.thomas02@imperial.ac.uk; p.naylor@imperial.ac.uk).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TASL.2009.2022430
1558-7916/$26.00 © 2009 IEEE
voice source modeling [6]. Identification of GOIs is necessary
for closed-phase LPC analysis and subsequent inverse filtering
to obtain an estimate of glottal volume flow from a speech
signal [7] for applications such as feature extraction for speaker
identification [8]. Further uses are found in the analysis of
pathological speech, including types of dysphonia [9], vocal
fold impact stress [10] and essential tremor [11]. We refer to
GOIs and GCIs as glottal activity.
The Electroglottograph (EGG) (or Laryngograph) signal
[12] is a measurement of the electrical conductance of the
glottis captured contemporaneously with speech recordings.
The measured EGG signal is proportional to the glottal contact
area, whose derivative (DEGG) during voiced speech contains
short, high-amplitude impulsive temporal features (spikes) due
to glottal closure and smaller features of opposite sign due to
glottal opening. An example of a voiced speech segment, the
corresponding EGG recording and its derivative is shown in
Fig. 1. Many approaches analyze the EGG by searching for
spikes in DEGG [13]–[16] and compare their amplitudes with
thresholds to obtain an estimate of glottal activity during voiced

Authorized licensed use limited to: Imperial College London. Downloaded on January 4, 2010 at 08:01 from IEEE Xplore. Restrictions apply.
1558 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 8, NOVEMBER 2009
speech. Recent approaches have applied multiscale analysis
to detect glottal activity as singularities in the EGG signal
[17] and speech signal [18]. Existing techniques are, however,
often prone to errors around the end of voicing as discussed in
Section II.
In cases where only the speech signal is available, new al-
gorithms have recently been proposed which estimate glottal
activity from the speech signal alone [19]–[23], and this is an
ongoing topic of research with seemingly ever-improving re-
sults. These algorithms enable glottal activity information to be
determined in real-world applications in which, typically, the
EGG signal is not available. However, as such methods improve,
their evaluation requires ever more accurate references. This re-
quirement, alongside the application to the study of pathological
speech, further motivates the development of better EGG-based
detection algorithms.
This paper describes the Singularity in EGG by Multiscale
Analysis (SIGMA) algorithm. SIGMA benefits from the use
of multiscale processing but it novely extends the approach by
performing spike detection on the multiscale product using a
group delay method [24] which circumvents the need for thresh-
olding. The robustness of our approach to false detections is
further enhanced by Gaussian mixture modeling [25] which is
used to remove detections with unlikely features. The proposed
method provides GCI estimation with outstandingly high accu-
racy which also achieves similarly accurate GOI detection. Ad-
ditionally, the algorithm makes no assumptions about the nature
of the EGG signal other than the bounds on the range of glottal
frequency and open quotients [26]; SIGMA may therefore have
many further uses as it is also suitable for singularity detection
in applications outside the field of speech processing.
This paper is organized as follows. Section II reviews the
characteristics of the EGG signal and the methodology em-
ployed by some existing algorithms. Section III describes
multiscale analysis, the use of the group delay function and
Gaussian mixture modeling for spike detection in the multiscale
product. The proposed SIGMA algorithm is compared with
existing techniques and evaluated in Section IV. Conclusions
are drawn in Section V.
II. INTERPRETING EGG SIGNALS
A. HQTx and TXGEN
In Section IV, the performance of SIGMA is compared
with two existing algorithms: High Quality Time of excitation
(HQTx) and Time of eXcitation GENerator (TXGEN) [16].
The following is a brief description of their operation.
HQTx uses two derived functions: DEGG and an estimation
of instantaneous gradient. A threshold function varies dynami-
cally with the EGG signal, whose minimum is set by periods of
silence assumed to lie during the first and last 20 ms of the EGG
recording. The instants of time when the DEGG and instanta-
neous gradient exceed this threshold are the estimated GCIs.
TXGEN uses a more straightforward approach but attempts
to detect both GCIs and GOIs. After low-pass filtering the EGG
signal at 3 kHz, it is differentiated to find DEGG. High and
low thresholds are set by the extrema of the DEGG signal from
Authorized licensed use limited to: Imperial College London. Downloaded on January 4, 2010 at 08:01 from IEEE Xplore. Restrictions apply.
the entire recording multiplied by constant-valued coefficient.
If DEGG passes through both thresholds within a set period of
time, an estimated GCI is flagged. A GOI is the point in the
EGG signal whose amplitude is equal to the amplitude at the
preceding GCI.
B. Detection Errors
Both SIGMA and the algorithms described are evaluated
against a large hand-labeled database. The remainder of this
section describes common features of the EGG signal, those
cases where interpretation of the EGG signal requires clarifica-
tion and the resulting errors made by existing algorithms.
A voiced speech signal, its corresponding time-aligned
EGG signal and the EGG derivative are shown in Fig. 1. Time
alignment is achieved by ensuring that the lip-microphone
propagation distance plus an estimate of the length of the
talker’s vocal tract is a constant value, then subtracting the
corresponding delay. We define a positive EGG signal to be
high glottal contact area, giving positive- and negative-going
transients for GCIs and GOIs, respectively, with corresponding
spikes in the EGG derivative.
Errors in GCI detection can be divided into two categories
[19]: False alarm errors are made when more than one GCI is
detected within a reference cycle; Miss errors are made when no
GCI is detected within a reference cycle (GOI errors are treated
in the same manner). Errors occur when certain types of EGG
signal, discussed in the following sections, cause a poor estimate
of the signal thresholds described in Section II-A.
C. “False Alarm” Errors
It has been shown that, for normal “modal” voiced speech,
the frequency of oscillation of the glottis and the open quo-
tient are dependent on phoneme and voice quality [12], [14].
Studies have further revealed that, for a given talker, the diffi-
culty of detecting glottal closure is largely independent of the
sound produced but that interesting effects occur at the bound-
aries of voiced/unvoiced speech, noting in particular [27]:
1) “Vocal fold vibration does not stop abruptly at the end of
voicing, but slowly decays as the vocal folds come to a rest
position.”
2) “It is possible for vocal fold vibration to continue without
the generation of any significant energy,” termed “breathy
offsets” [28].
This is examined in greater detail in [28], where a third phenom-
enon is observed at the end of voicing.
1) “A persistence of energy in the speech waveform after
the EGG waveform has dropped virtually to zero,” termed
“breathy voice.”
In the case of breathy offsets, GCIs can be detected from the
EGG long after the speech amplitude has significantly dimin-
ished as the EGG signal remains modal, with increasing open
phases that result in a breathier sound [28]. This is demonstrated
in Fig. 2, showing 14 cycles of breathy offset terminating in
breathy voice when EGG signal finally loses modality.
In the case of breathy voice, observed throughout case (3)
and at the very end of case (2), the glottis is “flapping in the
breeze” [29] with insufficient contact to register on the EGG
waveform. As described in [30], “If the glottis does not shut
THOMAS AND NAYLOR: SIGMA ALGORITHM: A GLOTTAL ACTIVITY DETECTOR FOR EGG SIGNALS
Fig. 2. (a) Speech signal, (b) EGG signal, and (c) its time derivative with over-
layed HQTx GCI estimation markers at the end of a voiced speech segment, /u/,
exhibiting “breathy offset” (cycles 8–21) and briefly “breathy voice.” The first
22 GCIs are identified correctly (marked “ ”) but the last three (marked “ ”)
are erroneous.
Fig. 3. (a) Speech signal, (b) EGG signal, and (c) its time derivative with over-
layed HQTx GCI estimation markers at the end of a voiced speech segment, /I/,
exhibiting “breathy voice.” The first three GCIs are identified correctly (marked
“ ”) but the last four (marked “ ”) are erroneous. Negative peaks due to glottal
opening are significant in (c).
quickly enough no vocal wave is generated in the supra-
glottic cavity,” and is demonstrated in Fig. 3. In both cases,
a number of erroneous GCIs are detected by HQTx during
segments of breathy voice ( ) until its dynamic threshold is
no longer exceeded. These errors also often occur at erratic
intervals. For the hand-labeled reference, marked “ ,” the
labeler would not mark any GCIs where there is no visible
instant defining the periodicity, as would be the case with all
instances of breathy voice.
Breathy voice represents a natural transition from modal
voiced speech to unvoiced or silence [28]. It is further noted
Authorized licensed use limited to: Imperial College London. Downloaded on January 4, 2010 at 08:01 from IEEE Xplore. Restrictions apply.
1559
Fig. 4. (a) Original Speech signal with correct GCIs (marked “ ”) and false
alarm errors (marked “ ”) and (b) time scale expanded by three times with
the PSOLA Algorithm. Voiced cycles are copied and concatenated to increase
duration; this works well for modal speech but fails when GCIs are detected in
the wrong location.
that this usually lasts for just a few cycles of speech but er-
roneous estimates by a GCI detector during these segments
can cause significant problems for glottal-synchronous algo-
rithms. For example, a pitch tracker [2] that calculates pitch
on a cycle-by-cycle basis will give highly erratic results.
Glottal-synchronous speech processing algorithms such as
prosodic speech modification [3], speech dereverberation [4],
speech synthesis [5], and voice source modeling [6] all rely
upon the manipulation of individual cycles of speech. Any
fricatives or plosives following segments of voiced speech
will be treated as periodic, giving rise to particularly annoying
artefacts [31].
An example is shown in Fig. 4 where HQTx is used to drive
the PSOLA algorithm [3] to increase the duration of a speech
signal by three times without affecting prosody or formant struc-
ture. Applications for increasing the duration of a speech signal
include enhancing intelligibility and lip synchronization in mo-
tion video. It is achieved by repeating cycles of voiced speech
and concatenating them with an estimate of the correct period
as shown in the first 70 ms of Fig. 4(b). Unvoiced speech and
voiced-unvoiced transitions do not exhibit such periodicity so a
common approach is to leave these segments unmodified [31].
This is not the case due to the erroneous detections at the voiced-
unvoiced transition from 70–150 ms, leading to strange artefacts
that detract from the otherwise natural sound of the processed
voiced speech segments.
Sudden changes in EGG amplitude can also cause false alarm
errors in dynamic threshold-based algorithms if the threshold is
too low. A further problem with dynamic thresholds arises when
GCIs have slow rise times [17], causing not a spike but a spread
pulse in the DEGG. In this case, we define the GCI as the center
of energy of the pulse.
D. “Miss” Errors
A common feature at the end of voiced segments is a reduced
EGG signal amplitude compared with normal modal voice.
1560 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 8, NOVEMBER 2009
TXGEN’s thresholds are proportional to the extrema of the en-
tire signal and it is generally not prone to the false alarm errors
exhibited by HQTx. It instead gives miss errors where the EGG
amplitude is consistently low, particularly at the very beginning
and very end of voiced speech segments. For the majority of
glottal-synchronous algorithms this does not pose a significant
problem. If, however, the amplitude of the EGG signal mo-
mentarily drops below the fixed threshold, TXGEN can miss
a small number of isolated cycles which can be problematic
for certain applications. Data-Driven Voice Source Modeling
[6], for example, derives feature vectors from individual of
cycles of voiced speech which are then clustered to determine
classes of voice source. This has been demonstrated to have
applications in speech compression [6] and artificial bandwidth
extension [32]. A missed GCI results in features being derived
from multiple cycles of speech, causing misclassification and
distorting the processed signal.
HQTx can exhibit miss errors following a sudden decrease in
EGG amplitude due to smoothing of the dynamic threshold that
is not employed in TXGEN.
E. False Alarm/Miss Tradeoff
In general, HQTx is prone to false alarm errors, particularly
at the end of voiced segments. This is verified in Section IV; it is
further shown that miss errors are far less common. In contrast,
TXGEN is generally prone to miss errors with relatively few
false alarms; this is also verified in Section IV.
HQTx fails largely because thresholds are estimated over too
short a window and TXGEN because thresholds are based upon
single global thresholds for the whole speech utterance. The
constant of proportionality used to set the threshold from signal
extrema can be varied in TXGEN’s function call. The default
was empirically chosen to give the best tradeoff between miss
and false alarm errors; a marginally lower value can result in
increased false alarms and decreased misses. There is there-
fore a clear tradeoff between false alarms and misses caused
by the thresholding approach employed by the majority of ex-
isting algorithms. The severity of this type of error is applica-
tion-specific but, when used as a reference to evaluate speech-
based GCI/GOI detectors, neither should be deemed acceptable.
SIGMA instead employs a novel method for detecting GCIs and
GOIs that does not use thresholding, circumventing the false
alarm/miss tradeoff and providing accurate estimates for the en-
tire EGG signal.
F. EGG at Glottal Opening
A glottal closure instant is usually followed by a GOI, which
manifests itself as a weaker spike of opposite sign in the EGG
derivative [13] whose amplitude is largely speaker-dependent.
GOI detection suffers from the same problems as GCI detec-
tion but is more challenging because of the low amplitude of
the opening pulses. Compare the negative halves of the EGG
signals in Figs. 1 and 3. In Fig. 1, glottal opening results in
spread pulses, hence the concept of an opening phase rather
than opening instant is often used. However, as in the case of
a spread-pulse GCI, we consistently define the GOI as its center
of energy. Fig. 3 represents a speaker for whom the opening
spikes in DEGG are easier to locate.
Authorized licensed use limited to: Imperial College London. Downloaded on January 4, 2010 at 08:01 from IEEE Xplore. Restrictions apply.
Fig. 5. Three-level dyadic signal decomposition on a signal into detail,
, and approximation, , signals. (a) is the Dyadic Wavelet Transform
(DWT), and (b) the Stationary Wavelet Transform (SWT), an overcomplete ver-
sion of the DWT useful in the detection of discontinuities.
III. GLOTTAL ACTIVITY DETECTION WITH
THE SIGMA ALGORITHM
Detection of glottal activity from an EGG signal involves iso-
lating regions of discontinuity, sometimes referred to as singu-
larities. A common approach is the detection of spikes in the
derivative of the EGG signal, whose estimates are refined using
the peak amplitude of the DEGG and a longer-term measure of
the change in EGG amplitude as a cost function.
A. Multiscale Analysis
Let us consider a generalization of the HQTx approach that
employs two estimates of signal gradient. The dyadic wavelet
transform [33] involves iteratively decomposing a signal
into decimated subbands; a three-level decomposition is shown
in Fig. 5(a), where the downsampling and filtering operations
split the signal into octave-wide subbands.
The filters and have high- and low-pass character-
istics, respectively. It is shown in [34] that, for singularity detec-
tion in EGG signals, each filter in the filterbank should be a first-
order differentiation operator at increasing levels of smoothing.
A wavelet fulfilling this criteria is described as having one van-
ishing moment and discontinuities in the input signal are seen
as converging maxima across scales [35].
A derivative-of-Gaussian (dG) approximation with cubic
spline wavelet decomposition filters is used in [36] and [34]
which provides the differentiation and smoothing we require.
However, an arbitrary number of filters exist which fulfil the
same criteria. A number of derivations can be found in [37] but
give little idea as to their use in the detection of singularities.
In order to determine the relative performance, the proposed
algorithm was run with five different sets of decomposition fil-
ters. Section IV-C presents a performance comparison between
the chosen wavelet, whose filters are shown in Fig. 6, and the
popular cubic spline dG wavelet.
The dyadic wavelet transform is dyadic in both scale and
time. Only scale is of interest in singularity detection, so we
do not decimate as shown in Fig. 5(b). Instead, the filters
and are upsampled by 2 at each iteration to implement the
THOMAS AND NAYLOR: SIGMA ALGORITHM: A GLOTTAL ACTIVITY DETECTOR FOR EGG SIGNALS 1561

Fig. 6. (a) Approximation and (b) detail analysis filters for multiscale analysis. Fig. 7. EGG waveform, multiscale product and group delay function for GCI
Iterating these filters through a dyadic filterbank constructs a biorthogonal spline detection. Candidates are marked “ ” and chosen candidates are marked “ .”
wavelet with one vanishing moment. The ideal slope, marked in a dashed line on the lowest plot, is the slope which
would exist if the candidates were perfect impulses.

change of scale to form and at scale . This over-

complete representation of a signal is discussed in detail in [35] becomes prohibitively large, demanding high processing
and is given many names including: Stationary Wavelet Trans- resources and smoothing adjacent discontinuties. is
form (SWT), Algorithme à Trous (Hole Algorithm), Redundant deemed a good compromise [36].
Wavelet Transform (RWT) and Undecimated Wavelet Transform B. Group Delay Function
(UWT). The signal’s length remains unchanged throughout the
filterbank tree, allowing simple sample-by-sample multiplica- A group delay function (GD) [24] can be used for detection
tion of the signal at different scales to find converging maxima. of peaks in linear prediction residuals of speech and can be ap-
Denote the wavelet , where plied to locate spikes in any signal if their minimum separation,
. The SWT of the EGG signal at scale is , is known. Consider the multiscale product, , and an
-sample windowed segment beginning at sample
(1)
(4)
where , plus the remaining coarse scale information
The group delay of is given by [38]
denoted . This is a simple linear filtering operation

(2) (5)

where is the SWT of at scale and are the ap- where is the Fourier transform of and is
proximation coefficients at scale . The multiscale product, the Fourier transform of at frequency . If
, is formed by , where is a unit impulse function, it
follows from (5) that . In the presence of noise,
remains constant but with a degree of additive noise, so
(3) an averaging procedure needs to be performed over ; different
approaches are reviewed in [24]. The Energy-Weighted Group
where it is assumed that the lowest scale to include is always 1. Delay was deemed the most appropriate [20], defined as
The sign of is inverted compared with a DEGG using the
chosen wavelet, hence a minus sign is included to maintain the
convention. The de-noising effect of at each scale in con- (6)
junction with the multiscale product means that is near-
zero except at discontinuities across the first scales of
as depicted in Fig. 7(b), allowing better identification of discon- Manipulation yields the simplified expression
tinuties than the DEGG. The function can be half-wave
rectified to contain peaks pertaining only to GCIs, , or
GOIs, , which aids the group delay function in the fol- (7)
lowing step. The value of is limited by , but it is often no
greater than as the region of support (RoS) of and

Authorized licensed use limited to: Imperial College London. Downloaded on January 4, 2010 at 08:01 from IEEE Xplore. Restrictions apply.
1562 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 8, NOVEMBER 2009

which is an efficient time-domain formulation and can be

viewed as the “center of energy” of , bounded in the
range . The location of the nega-
tive-going zero crossings of give an accurate estimation
of the location of a spike in a function as depicted in Fig. 7(c).
Additionally, if a spike is spread in time then the group delay
method will find its center of energy, which is particularly
useful in the case of the “redoubled” GCI discussed in [17].
The same analysis is applied to to provide , whose
negative-going zero crossings are GOI candidates.

C. Candidate Selection
The true GCIs are usually a subset of the negative-going zero
crossings of , with additional false crossings during un-
voiced speech, silence and occasionally between GCIs. Many
existing approaches concentrate only on those areas where false Fig. 8. Typical distribution of GCI feature vectors for a segment of voiced/
candidates are unlikely to occur. The following candidate selec- unvoiced/silent speech. The chosen cluster, whose members are marked “ ” is
tion technique aims to remove all false candidates to provide a the one whose mean is furthest from the origin. Rejected candidates are
marked “ .”
set of true GCIs throughout an entire segment of speech. Let the
number of candidates be occurring at samples ,
. Three measurements construct a
upon the likelihood of class , , given feature vector
feature vector, , from which is derived
a feature matrix, . The features are
defined as follows. (11)
1) Consistency of the group delay gradient. In the case of
a Dirac pulse, is a negative unit slope, with a Fig. 8 shows a typical distribution of the feature vectors for
zero crossing at the location of the impulse and width a segment of mixed voiced/unvoiced/silent speech. It has been
samples, as shown in Fig. 7(c). A spread pulse or the found empirically that the cluster whose mean is furthest
presence of noise will cause the slope to deviate from from the origin is most likely to contain the chosen candidates,
the ideal shape, denoted . The RMS error between marked “ .” Rejected candidates are marked “ .” The chosen
ideal and measured is calculated as GCI estimates are defined as .
GOIs are calculated in the same way but with reversed signs
where appropriate.
(8)
D. Swallowing
The algorithm proposed thus far performs accurate singu-
2) Peak value of multiscale product’s root inside group larity detection on an input signal without considering any
delay window. It is shown in [34] that the root of characteristics peculiar to EGG waveforms. It is found that
helps to give a “zooming in” on the signal, par- in natural conversional speech, singularities are often caused
ticularly at weak amplitudes (in this case ). Exper- by swallowing and occasionally by electrical interference
imentation with this algorithm has shown that the group in the measurement apparatus and are usually single iso-
delay function gives best results on but that its lated impulse-like signals. Considering a maximum period
root has better discriminative properties. all GCIs which are separated from a neighbouring
GCI by more than are rejected, else they are kept pro-
viding: .
(9) Experimentation has shown that provided the polarity of the
3) Area beneath multiscale product’s root inside group recording is correct, swallowing only causes errors in closure
delay window. In the case of a spread singularity, the detection so this technique is not applied to opening detection.
area beneath the multiscale product’s root can pro-
vide better discrimination of candidates. E. GOI Postfiltering
GOIs are detected from using the same approach
(10) as applied to GCI detection (with inverted signs where appro-
priate). However, the energy imparted by glottal opening is often
significantly lower than glottal closure, which results in more er-
The distributions of the feature vectors are modeled as two roneous GOI candidates. Assuming that a GOI always accompa-
multivariate Gaussians using the EM algorithm [25], initialized nies a GCI, postprocessing can be applied to use GCI estimates
with two random data points. Acceptance or rejection is based to improve GOIs accordingly.

Authorized licensed use limited to: Imperial College London. Downloaded on January 4, 2010 at 08:01 from IEEE Xplore. Restrictions apply.
THOMAS AND NAYLOR: SIGMA ALGORITHM: A GLOTTAL ACTIVITY DETECTOR FOR EGG SIGNALS 1563

Fig. 9. SIGMA system diagram. The EGG signal is decomposed into multiple scales from which the half-wave rectified multiscale product is derived.
Spike detection is performed on by the negative-going zero crossings of the group delay function at samples . Feature vectors derived from
the ideal group delay slope and are clustered by an unsupervised EM algorithm to obtain the GCI estimates . Similarly, GOIs are detected using
the negative half-wave of the multiscale product . Postprocessing is applied to the GCI estimates to remove isolated clicks from sources other than glottal
closure to give . GOI postprocessing removes candidates which do not lie within the range of permitted open quotients, using the GCIs as references giving .

The main cause of error in GOI post-filtering is small pertur-
bations in immediately preceding a glottal closure which
triggers a zero crossing in the group delay function. A region
surrounding the closure is therefore isolated, limiting the al-
lowed open quotient, , to the bounds and . The
first candidate which lies within these limits is accepted; if no
candidate is found, then one is inserted following the current
GCI at the previous open quotient.
The SIGMA system diagram is shown in Fig. 9. Symmetry
can be seen between closure and opening detection up until the
postprocessing stage; prior to this point the algorithm need only
know the maximum frequency of the singularities to detect and
so is suitable for general singularity detection.
IV. RESULTS AND DISCUSSION
The SIGMA algorithm has three parameters and these were
set as follows.
• : the group delay evaluation window size and there-
fore the maximum frequency of singularities which can be
detected. In the case of voiced speech, the maximum glottal
frequency is Hz giving ms.
• : the maximum glottal period, so that isolated GCI
candidates separated from neighboring candidates by more
than this value are removed in the GCI postfiltering step.
A minimum glottal frequency of 50 Hz leads to a
ms.
• : the minimum and maximum open quotients
for GOI postfiltering. Their purpose is to isolate a region
around a GCI inside which a GOI cannot be detected. They
are set at 10% and 90%, respectively.
The MATLAB implementation of the chosen biorthogonal
spline decomposition filters is called bior1.5.
A. Experiment 1: Evaluation With APLAWD and SAM
The APLAWD database [39] contains speech and contem-
poraneous EGG recordings of five short sentences, repeated
ten times by five male and five female talkers. GCIs and
GOIs were hand-labeled on the first repetition of every
sentence independently of the algorithms under test, de-
noted , , and , SIGMA’s miss, FAT and FA measures are broadly similar which
, respectively. A subset of the
SAM database [40] contains readings of duration approxi-
mately 150 seconds by two male and two female speakers
and these were labeled in the same manner. SAM recordings
are considered to contain more natural speech with a greater
number of swallows and present a more challenging task for a
glottal activity detector. The EGG recordings were run through
the HQTx (GCI only), TXGEN and SIGMA algorithms and
were evaluated by finding the number of estimates per reference
cycle then classified as follows, depicted in Fig. 10.
1) Hit. One estimate per true glottal cycle.
2) Miss. No estimates per true glottal cycle.
3) False Alarm (FA). More than one estimate per glottal cycle.
4) False Alarm Total (FAT) Total number of false alarms (the
number of estimates which are not hits).
The measures are defined as follows.
1) .
2) .
3) .
4) .
5) .
A glottal cycle is defined as for GCIs and
for GOIs. Hit accuracy and hit bias
are the RMS and mean errors between all hits and the cor-
responding ground-truth estimates, respectively. The testing
strategy is identical to that employed in [19] with the addition
of the FAT measure, which counts the total number of false
alarms as a proportion of total estimates and not the number
of reference cycles containing more than one estimate as a
proportion of true glottal cycles. The overall figure of merit
provides a single-valued measure of performance by expressing
the hit rate as a proportion of all reference cycles summed with
the number of non-hit estimates (the FAT).
The GCI results in Tables I and III show that SIGMA per-
forms significantly better than HQTx and TXGEN when applied
to either database. Notably HQTx is prone to false alarm errors
whereas TXGEN is prone to miss errors; this agrees with the
qualitative analysis of HQTx’s performance in Section II which
showed that it is prone to false alarms at the end of segments of
voiced speech. HQTx and TXGEN exhibit much greater FAT
than FA which suggests that each false alarm is usually fol-
lowed by successive false alarms within a single reference cycle.
tells us that successive false alarms do not usually occur within a

Authorized licensed use limited to: Imperial College London. Downloaded on January 4, 2010 at 08:01 from IEEE Xplore. Restrictions apply.
1564 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 8, NOVEMBER 2009
Fig. 10. Testing strategy. A hit is one estimate occuring during a reference
cycle. A miss is the absence of an estimate per reference cycle. If more than one
estimate occurs per reference cycle, one false alarm (FA) is counted and the total
number false alarms in the cycle are added to false alarm total (FAT). Accuracy
and bias are the RMS and mean errors between hits and the corresponding ref-
erence, respectively.
TABLE I
CLOSURE PERFORMANCE ON THE APLAWD DATABASE BY HQTX, TXGEN,
SIGMA (dG AND bior1.5 WAVELET) ALGORITHMS
TABLE II
OPENING PERFORMANCE ON THE APLAWD DATABASE BY TXGEN, SIGMA
(dG AND bior1.5 WAVELET) ALGORITHMS
TABLE III
CLOSURE PERFORMANCE ON THE SAM DATABASE BY HQTX, TXGEN,
SIGMA (dG AND bior1.5 WAVELET) ALGORITHMS
given reference cycle and that misses and false alarms have sim-
ilar likelihood. SIGMA’s overall figures of merit are more than
an order of magnitude greater than the other algorithms under
test.
SIGMA’s GCI hit accuracy is in the order of a few samples
which agrees with the statement in Section III-C that the true
GCIs are usually a subset of the SIGMA candidate GCIs before
Authorized licensed use limited to: Imperial College London. Downloaded on January 4, 2010 at 08:01 from IEEE Xplore. Restrictions apply.
TABLE IV
OPENING PERFORMANCE ON THE SAM DATABASE BY TXGEN, SIGMA (dG
WAVELET AND bior1.5 WAVELET) ALGORITHMS
clustering. SIGMA and HQTx hit bias are universally low but
TXGEN’s estimates tend to occur slightly early.
SIGMA’s GOI results in Tables II and IV are also encour-
aging. The reliance upon the estimated GCIs results in similar
hit, miss and false alarm rates, with diminished hit accuracy due
to the greater difficulty of precisely locating openings. The gap
in the overall figure of merit between SIMGA and TXGEN is
again more than an order of magnitude.
B. Experiment 2: Variation in Group Delay Window Size
The group delay evaluation window size was set according to
the physical constraints of human speech, whose minimum fun-
damental period is around 2.5 ms. This experiment assesses the
algorithm’s sensitivity to variation in the group delay window
size on the APLAWD database.
The results presented in Fig. 11 show that 2.5 ms is indeed
an optimal choice of window length. The reliance on GCIs to
estimate GOIs means that intuitively the overall, hit, miss, and
FAT rates should vary in a similar manner which is confirmed
by these results.
FAT rates increase with decreasing window sizes due to the
fact that more negative zero crossings can occur in the group
delay function per unit time. In this case the true candidates re-
main a subset of all candidates, with a number of additional false
ones arising. Providing the clustering algorithm can discrimi-
nate against the false candidates, those which are true should
always be detected so false alarm rates should therefore increase
slowly with decreasing window size.
Miss rates increase with window size as neighboring singular-
ities can occur within a single group delay window and reduce
the number of negative zero crossings. It becomes impossible
for the GMM to find the correct candidates as they are no longer
a subset of the candidate set, hence miss rates climb rapidly with
increasing window size.
GCI bias and hit accuracy are relatively immune to variations
in window size, suggesting that providing one candidate occurs
per true period, is it statistically the correct choice. GOI bias and
hit accuracy are more sensitive, showing the most significant in-
crease with reduced window size. Bias increases monotonically
with decreasing window length.
This experiment was repeated for male- and female-only
speech. The results provide similar curves to the previous ex-
periment that employs both genders, the optimum value being
shifted up to approximately 3 ms for male voices and down to
approximately 2 ms for female. The experiment with mixed
male/female speech shows that variation in group delay size
does not have a significant effect upon the results in the range
of approximately 1.5 to 3.5 ms, hence performance is weakly
dependent on gender.
THOMAS AND NAYLOR: SIGMA ALGORITHM: A GLOTTAL ACTIVITY DETECTOR FOR EGG SIGNALS 1565

correct estimates are removed by the clustering of three-dimen-

sional feature vectors using the EM algorithm. Postprocessing
removes isolated GCIs and uses GCIs to aid GOI detection.
A comparison was made between the proposed approach and
two popular existing methods by evaluating their performance
against 50 short and four long hand-labeled sentences. An ex-
isting testing procedure with some new enhancements was used,
showing very accurate GCI and GOI detection with the pro-
posed method, fulfilling our objective of obtaining results that
are accurate enough to be used as a reference. Our method en-
ables accurate evaluation of speech-based glottal activity detec-
tion algorithms, precise estimation of the closed phase for the
estimation of glottal volume flow and could also be applied to
the analysis of a number of types of pathological speech. Fur-
ther, few assumptions are made about the nature of the input
signal. This allows the application of the proposed algorithm
to singularity detection in almost any signal provided the min-
imum separation of singularities is known.

Fig. 11. Effect of varied group delay window length on (a) overall and hit, (b)
miss and FAT, and (c) and bias and hit accuracy. The choice of 2.5 ms from
physical reasoning is close to the optimal value.
C. Experiment 3: Comparison With Cubic Spline Wavelet
The derivative-of-Gaussian (dG) cubic spline wavelet is the
wavelet of choice for multiscale analysis in [17], [18] and [36].
Experiments with other common wavelets have shown that the
bior1.5 biorthogonal spline wavelet is more effective for EGG
analysis with this algorithm. The results in Tables I–IV show
SIGMA using the dG wavelet (labeled SIGMA-dG) as well as
the proposed bior1.5 (labeled SIGMA).
The performance of SIGMA is slightly reduced with the dG
wavelet, particularly with increased false alarms and increased
hit error on the opening tests. Miss rates are slightly reduced but
the greater increase in false alarm rate diminishes the overall
performance results.
V. CONCLUSION
We have shown that robust detection of GCIs and GOIs from
EGG signals is particularly challenging at the transition regions
around the ending of voicing. A new method for glottal activity
detection from EGG recordings has been presented which is
accurate even in these challenging regions. It first detects sin-
gularities in the EGG signal by the multiscale product of three
dyadic scales. It then employs a technique based upon the group
delay function which detects peaks in the multiscale product. In-
REFERENCES
[1] P. Davies, G. A. Lindsey, H. Fuller, and A. J. Fourcin, “Variation of
glottal open and closed phases for speakers of English,” Proc. Inst.
Acoust., vol. 8, no. 7, pp. 539–546, 1986.
[2] W. Hess and H. Indefrey, “Accurate pitch determination of speech sig-
nals by means of a laryngograph,” in Proc. IEEE Intl. Conf. Acoust.,
Speech, Signal Process. (ICASSP), 1984, vol. 9, pp. 73–76.
[3] H. Valbret, E. Moulines, and J. P. Tubach, “Voice transformation using
PSOLA technique,” Speech Commun., vol. 11, no. 2, pp. 175–187, Jun.
1992.
[4] N. D. Gaubitch, P. A. Naylor, and D. B. Ward, “Multi-microphone
speech dereverberation using spatio-temporal averaging,” in Proc. Eur.
Signal Process. Conf. (EUSIPCO), Vienna, Austria, Sep. 2004, pp.
809–812.
[5] E. Moulines and F. Charpentier, “Pitch-synchronous waveform
processing techniques for text-to-speech synthesis using diphones,”
Speech Commun., vol. 9, no. 5–6, pp. 453–467, Dec. 1990.
[6] M. R. P. Thomas, J. Gudnason, and P. A. Naylor, “Data-driven voice
source waveform modelling,” in Proc. IEEE Intl. Conf. Acoust.,
Speech, Signal Process. (ICASSP), Taipei, Taiwan, Apr. 2009, pp.
3965–3968.
[7] J. Deller, “Some notes on closed phase glottal inverse filtering,” IEEE
Trans. Acoust., Speech, Signal Process., vol. ASSP–29, no. 4, pp.
917–919, Aug. 1981.
[8] J. Gudnason and M. Brookes, “Voice Source cepstrum coefficients
for speaker identification,” in Proc. IEEE Intl. Conf. Acoust., Speech,
Signal Process. (ICASSP), 2008, pp. 4821–4824.
[9] R. Colton and J. Casper, Understanding Voice Problems: A Physiolog-
ical Perspective for Diagnosis and Treatment. New York: Williams
& Wilkins, 1996.
[10] K. Verdolini, R. Chan, I. R. Titze, M. Hess, and W. Bierhals, “Corre-
spondence of electroglottographic closed quotient to vocal fold impact
stress in excised canine larynges,” J. Voice, vol. 12, no. 4, pp. 415–423,
Feb. 1998.
[11] J. Gamboa, F. J. Jiménez-Jiménez, A. Nieto, I. Cobeta, A. Vegas, M.
Orti-Pareka, T. Gasalla, J. A. Molina, and E. Garcia-Albea, “Acoustic
voice analysis in patients with essential tremor,” J. Voice, vol. 12, no.
4, pp. 444–452, Feb. 1998.
[12] E. R. M. Abberton, D. M. Howard, and A. J. Fourcin, “Laryngographic
assessment of normal voice: A tutorial,” Clinical Linguist. Phon., vol.
3, pp. 281–296, 1989.
[13] D. G. Childers, D. M. Hooks, G. P. Moore, L. Eskenazi, and A. L.
Lalwani, “Electroglottography and vocal fold physiology,” J. Speech.
Hear. Res., vol. 33, no. 2, pp. 245–254, Jun. 1990.
[14] D. M. Howard, “Variation of electrolaryngographically derived closed
quotient for trained and untrained adult female singers,” J. Voice, vol.
9, no. 2, pp. 121–1223, Jun. 1995.
[15] N. Henrich, C. d’Alessandro, M. Castellengo, and B. Doval, “On the
use of the derivative of electroglottographic signals for characterization
of nonpathological voice phonation,” J. Acoust. Soc. Amer., vol. 115,
no. 3, pp. 1321–1332, Mar. 2004.

Authorized licensed use limited to: Imperial College London. Downloaded on January 4, 2010 at 08:01 from IEEE Xplore. Restrictions apply.
1566 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 8, NOVEMBER 2009
[16] M. A. Huckvale, “Speech Filing System: Tools for Speech,” Tech. Rep.
Univ. College London, London, U.K., 2004 [Online]. Available: http://
www.phon.ucl.ac.uk/resource/sfs
[17] A. Bouzid and N. Ellouze, “Multiscale product of electroglottogram
signal for glottal closure and opening instant detection,” in Proc.
IMACS MultiConf. Comput. Eng. Syst. Applicat., 2006, vol. 1, pp.
106–109.
[18] A. Bouzid and N. Ellouze, “Glottal opening instant detection from
speech signal,” in Proc. Eur. Signal Process. Conf. (EUSIPCO), Vi-
enna, Austria, Sep. 2004, pp. 729–732.
[19] P. A. Naylor, A. Kounoudes, J. Gudnason, and M. Brookes, “Estima-
tion of glottal closure instants in voiced speech using the DYPSA algo-
rithm,” IEEE Trans. Speech Audio Process., vol. 15, no. 1, pp. 34–43,
Jan. 2007.
[20] M. R. P. Thomas, N. D. Gaubitch, and P. A. Naylor, “Multichannel
DYPSA for estimation of glottal closure instants in reverberant
speech,” in Proc. Eur. Signal Process. Conf. (EUSIPCO), Poznan,
Poland, Sep. 2007.
[21] K. S. Rao, S. R. M. Prasanna, and B. Yegnanarayana, “Determination
of instants of significant excitation in speech using Hilbert envelope
and group delay function,” IEEE Signal Process. Lett., vol. 14, no. 10,
pp. 762–765, Oct. 2007.
[22] K. S. R. Murty and B. Yegnanarayana, “Epoch extraction from speech
signals,” IEEE Trans. Audio, Speech, Lang. Process., vol. 16, no. 8, pp.
1602–1613, Nov. 2008.
[23] W. Saidi, A. Bouzid, and N. Ellouze, “Evaluation of multi-scale
product method and DYPSA algorithm for glottal closure instant de-
tection,” in Proc. 3rd Int. Conf. Inf. Commun. Technol.: From Theory
to Applicat. (ICTTA), Apr. 2008, pp. 1–5.
[24] M. Brookes, P. A. Naylor, and J. Gudnason, “A quantitative as-
sessment of group delay methods for identifying glottal closures in
voiced speech,” IEEE Trans. Speech Audio Process., vol. 14, no. 2,
pp. 456–466, Mar. 2006.
[25] A. P. Dempster, N. M. Laird, and D. B. Rubin, “Maximum likelihood
from incomplete data via the EM algorithm,” J. R. Statist. Soc., Ser. B,
vol. 39, no. 1, pp. 1–38, 1977.
[26] M. B. Higgins and J. H. Saxman, “A comparison of selected phonatory
behaviours of healthy aged and young adults,” J. Speech Hear. Res.,
vol. 34, pp. 1000–1010, Oct. 1991.
[27] A. K. Krishnamurthy and D. G. Childers, “Two-channel speech anal-
ysis,” IEEE Trans. Acoust., Speech, Signal Process., vol. ASSP-34, no.
4, pp. 730–743, Aug. 1986.
[28] D. M. Howard and G. Lindsey, “Conditioned variability in voicing off-
sets,” IEEE Trans. Acoust., Speech, Signal Process., vol. 36, no. 3, pp.
406–407, Mar. 1988.
[29] J. C. Catford, Fundamental Problems in Phonetics. Bloomington, IN:
Indiana Univ. Press, 1977.
[30] Y. Lebrun and J. Hasquin-Deleval, “On the so-called ’dissociations’ be-
tween electroglottogram and phonogram,” Folia Phoniatrica, vol. 23,
pp. 225–227, 1971.
[31] M. R. P. Thomas, J. Gudnason, and P. A. Naylor, “Application of the
DYPSA algorithm to segmented time-scale modification of speech,” in
Proc. Eur. Signal Process. Conf. (EUSIPCO), Lausanne, Switzerland,
Aug. 2008.
[32] P. Jax and P. Vary, “On artificial bandwidth extension of telephone
speech,” Signal Process., vol. 83, pp. 1707–1719, 2003.
[33] S. Mallat and S. Zhong, “Characterization of signals from multiscale
edges,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 14, no. 7, pp.
710–732, Jul. 1992.
Authorized
View publication stats licensed use limited to: Imperial College London. Downloaded on January 4, 2010 at 08:01 from IEEE Xplore. Restrictions apply.
[34] A. Bouzid and N. Ellouze, “Local regularity analysis at glottal opening
and closure instants in electroglottogram signal using wavelet trans-
form modulus maxima,” in Proc. Eurospeech, 2003, pp. 2837–2840.
[35] S. Mallat and W. L. Hwang, “Singularity detection and processing with
wavelets,” IEEE Trans. Inf. Theory, vol. 38, no. 2, pp. 617–643, Mar.
1992.
[36] B. M. Sadler and A. Swami, “Analysis of multiscale products for step
detection and estimation,” IEEE Trans. Inf. Theory, vol. 45, no. 3, pp.
1043–1051, Apr. 1999.
[37] I. Daubechies, Ten Lectures on Wavelets. Philadelphia, PA: SIAM,
1992.
[38] R. Smits and B. Yegnanarayana, “Determination of instants of signif-
icant excitation in speech using group delay function,” IEEE Trans.
Speech Audio Process., vol. 5, no. 3, pp. 325–333, Sep. 1995.
[39] G. Lindsey, A. Breen, and S. Nevard, “SPAR’s Archivable Actual-
Word Databases,” Tech. Rep. Univ. College London, London, U.K.,
1987.
[40] D. Chan, A. Fourcin, D. Gibbon, B. Granstrom, M. Huckvale,
G. Kokkinakis, K. Kvale, L. Lamel, B. Lindberg, A. Moreno,
J. Mouropoulos, F. Senia, I. Trancoso, C. Veld, and J. Zeiliger,
“EUROM – A spoken language resource for the EU,” in Proc. Eur.
Conf. Speech Commun. Technol., Sep. 1995, pp. 867–870.
Mark Thomas (S’06) received the M.Eng. degree in
electrical and electronic engineering from Imperial
College, London, U.K., in 2006 where he is currently
pursuing the Ph.D. degree.
His research interests include glottal-synchronous
and multichannel speech processing, involving
methods for analysis, prosodic manipulation and re-
verberation/noise reduction. His previous experience
in industry was with the BBC R&D Department,
where he worked on audio, video, and RF engi-
neering.
Patrick Naylor (M’89–SM’07) received the B.Eng.
degree in electronics and electrical engineering from
the University of Sheffield, Sheffield, U.K., in 1986
and the Ph.D. degree from Imperial College, London,
U.K., in 1990.
Since 1989, he has been a Member of Academic
Staff in the Communications and Signal Processing
Research Group, Imperial College London, where he
is also Director of Postgraduate Studies. His research
interests are in the areas of speech and audio signal
processing and he has worked in particular on adap-
tive signal processing for acoustic echo control, speaker identification, multi-
channel speech enhancement, and speech production modeling. In addition to
his academic research, he enjoys several fruitful links with industry in the U.K.,
U.S., and in mainland Europe.
Dr. Naylor is an Associate Editor of IEEE SIGNAL PROCESSING LETTERS and
a member of the IEEE Signal Processing Society Technical Committee on Audio
and Electroacoustics.